StockSense-AI / backend /optimizer_agent.py
PROSTLE's picture
feat: initial submission for StockSense-AI (clean history)
c165635
Raw History Blame Contribute Delete
5.52 kB
import os
import json
from feedback_engine import FeedbackEngine
from dynamic_params import DynamicParams, KEYS
# Re-using SDK logic from llm_analysis
GEMINI_AVAILABLE = False
try:
from google import genai as _genai_sdk
GEMINI_AVAILABLE = True
except ImportError:
pass
# Same model order as llm_analysis.py
GEMINI_MODELS = [
"gemini-2.5-flash",
"gemini-2.0-flash",
"gemini-2.0-flash-lite",
"gemini-flash-latest",
]
class OptimizerAgent:
def __init__(self):
self.api_key = os.getenv("GEMINI_API_KEY")
self.client = None
if GEMINI_AVAILABLE and self.api_key:
try:
self.client = _genai_sdk.Client(api_key=self.api_key)
except Exception as e:
print(f"[Optimizer] Init error: {e}")
def _call_llm(self, prompt: str) -> str:
"""Try each model in order, skipping on quota errors. Return raw text or raise."""
if not self.client:
raise RuntimeError("Gemini client not initialised β€” check GEMINI_API_KEY")
last_err = None
for model in GEMINI_MODELS:
try:
response = self.client.models.generate_content(
model=model, contents=prompt
)
print(f"[Optimizer] Used model: {model}")
return response.text.strip()
except Exception as e:
msg = str(e)
if "429" in msg or "RESOURCE_EXHAUSTED" in msg:
print(f"[Optimizer] {model}: quota exhausted, trying next...")
last_err = msg
continue
elif "404" in msg or "NOT_FOUND" in msg:
print(f"[Optimizer] {model}: not found, trying next...")
last_err = msg
continue
else:
raise
raise RuntimeError(f"All Gemini models quota-exhausted. Last error: {last_err}")
def run_optimization(self):
"""
Analyze recent performance and update dynamic parameters.
Falls back to a parameter-sanity review when there is no accuracy data yet.
Tries all Gemini model variants before giving up on quota errors.
"""
if not self.client:
return {"status": "error", "message": "LLM client unavailable β€” check GEMINI_API_KEY"}
# 1. Gather Data
report = FeedbackEngine.get_accuracy_report(days=7)
params = DynamicParams.get_all()
# 2. Build Optimization Prompt
# When no accuracy data exists yet, do a parameter-sanity warm-start.
if report:
perf_section = (
f"Current Performance Report (last 7 days):\n{json.dumps(report, indent=2)}\n\n"
"Based on the directional accuracy and average moves, suggest improvements:\n"
"- If Accuracy < 50%, consider increasing BUY_THRESHOLD.\n"
"- If XGB consistently outperforms LSTM, shift WEIGHT_XGB up and WEIGHT_LSTM down.\n"
"- If Stop Losses (ATR_MULTIPLIER_SL) hit too often, raise them slightly."
)
else:
perf_section = (
"No trade history yet β€” there is no accuracy data to analyse.\n"
"Please review the current parameters for sanity and suggest conservative\n"
"starting values appropriate for Indian mid-/large-cap intraday trading.\n"
"Typical safe ranges:\n"
" BUY_THRESHOLD: 60-72 | ATR_MULTIPLIER_SL: 1.2-1.8 | ATR_MULTIPLIER_TP1: 2.0-3.0\n"
" MIN_CONFIDENCE: 50-60 | WEIGHT_LSTM: 0.3-0.5 | WEIGHT_XGB: 0.3-0.5 | WEIGHT_SENTIMENT: 0.1-0.2"
)
prompt = f"""You are an AI ML Optimizer for a stock trading bot.
Current Dynamic Parameters:
{json.dumps(params, indent=2)}
{perf_section}
Respond ONLY with a JSON object of key-value pairs for the parameters you want to change.
Do NOT change parameters that are already at good values.
Example format: {{"BUY_THRESHOLD": 65, "ATR_MULTIPLIER_SL": 1.5}}"""
try:
# 3. Call LLM with multi-model fallback
text = self._call_llm(prompt)
if "```" in text:
text = text.split("```")[1]
if text.startswith("json"):
text = text[4:]
text = text.strip()
updates = json.loads(text)
# 4. Apply Updates (only for known keys)
applied = {}
for key, val in updates.items():
if key in params:
DynamicParams.update(key, val)
applied[key] = val
return {
"status": "success",
"updates_applied": applied,
"report_summary": report,
"note": "No trade history β€” parameter sanity review applied." if not report else ""
}
except RuntimeError as e:
# Quota or availability error β€” surface a clean message
msg = str(e)
print(f"[Optimizer] {msg}")
return {
"status": "quota_exhausted",
"message": "Gemini API daily quota exhausted. Optimization resumes automatically when quota resets (midnight Pacific Time).",
"raw": msg[:300]
}
except Exception as e:
print(f"[Optimizer] Optimization error: {e}")
return {"status": "error", "message": str(e)}